End-to-end HPC deployment automation: AzureHPC

This article will guide you how to automate the process to deploy an HPC environment on Azure, using an open source project AzureHPC.

The environment will look like below:

1. A Virtual Network named “hpcvnet" and a subnet named “compute".
2. A Virtual Machine named “headnode":
   • PBS job scheduler
   • 2TB NFS file system
3. A Virtual Machine Scale Set named “compute", which contains 2 instances.

Prerequisites:

1. A Linux environment, Windows 10 中的 WSL 2.0 as example with Azure CLI installed。Or you can just use CloudShell of Azure Portal.
2. An Azure subscription with sufficient quota including:
   • 1xDS8_v3 (8 cores)
   • 2xHC44rs (88 cores)

Step-by-step:

1. Download AzureHPC repo.

# log in to your Azure subscription
$ az login 

# mkdir your working environment
$ mkdir airlift 
$ cd airlift 

# Clone the AzureHPC repo 
$ git clone https://github.com/Azure/azurehpc.git
$ cd azurehpc 

# Source the install script 
$ source install.sh

There are many HPC templates in the /examples folder. Just change directory to the template folder and edit the config.json file.

2. Edit config.json.

# We will use /examples/simple_hpc_pbs template here
$ cd examples/simple_hpc_pbs 

# edit config.json
# vi config.json 
$ code .

Input values of “location", “resource_group", and “vm_type". You can also change network or storage settings as your desire.

3. Deploy the HPC environment.

$ azhpc-build

It should be completed around 10~15 mins.

4. Connect to “headnode" to check the HPC environment you just deployed.

Please note that /share is mounted, and /apps, /data, and /home directories are exported and accessible from PBS nodes. You can “ssh <PBSNODENAME>" to verify.

$ azhpc-connect -u hpcuser headnode
Fri Jun 28 09:18:04 UTC 2019 : logging in to headnode (via headnode6cfe86.westus2.cloudapp.azure.com)
$ pbsnodes -avS
$ df -h

5. As PBS is well installed and configured on “headnode", you can submit jobs and check jobs status.

$ qstat -Q

6. Delete the environment.

$ azhpc-destroy

Or just delete the whole resource group from Azure portal.